U.S. patent number 7,337,086 [Application Number 11/583,248] was granted by the patent office on 2008-02-26 for system and method for combining diagnostic evidences for turbine engine fault detection.
This patent grant is currently assigned to Honeywell International, Inc.. Invention is credited to Valerie Guralnik, Dinkar Mylaraswamy, Harold C. Voges.
United States Patent |
7,337,086 |
Guralnik , et al. |
February 26, 2008 |
System and method for combining diagnostic evidences for turbine
engine fault detection
Abstract
A system and method for combining conclusions from multiple
fault detection techniques to isolate likely faults in a turbine
engine is provided. The system and method provide the ability to
effectively deal with multiple concurrent faults in the engine.
Additionally, the embodiments of the invention provide the ability
to correctly characterize multiple conclusions generated from
evidence having different levels of interdependence. In one
embodiment, the conclusions based on device data with high
dependency are aggregated using a high dependency aggregation rule,
and the resulting high-dependency sets are then further aggregated
using a weak dependency rule. Finally, any conclusions based on
independent evidence can be aggregated using an independent
combination rule. The resulting aggregation determines which
fault(s) are most likely indicated by the plurality of conclusions,
taken into account the dependency of the device data used to
generate the conclusions.
Inventors: |
Guralnik; Valerie (Orono,
MN), Mylaraswamy; Dinkar (Fridley, MN), Voges; Harold
C. (Shoreview, MN) |
Assignee: |
Honeywell International, Inc.
(Morristown, NJ)
|
Family
ID: |
37949491 |
Appl.
No.: |
11/583,248 |
Filed: |
October 18, 2006 |
Prior Publication Data
|
|
|
|
Document
Identifier |
Publication Date |
|
US 20070088982 A1 |
Apr 19, 2007 |
|
Related U.S. Patent Documents
|
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
Issue Date |
|
|
60728088 |
Oct 18, 2005 |
|
|
|
|
Current U.S.
Class: |
702/113;
702/183 |
Current CPC
Class: |
G05B
23/0262 (20130101); G05B 23/0275 (20130101) |
Current International
Class: |
G01M
15/00 (20060101) |
Field of
Search: |
;702/113,188 |
References Cited
[Referenced By]
U.S. Patent Documents
Primary Examiner: Barlow; John
Assistant Examiner: Moffat; Jonathan
Attorney, Agent or Firm: Ingrassia Fisher & Lorenz
Parent Case Text
CROSS-REFERENCES TO RELATED APPLICATION
This application claims the benefit of U.S. Provisional Application
No. 60/728,088 filed Oct. 18, 2005.
Claims
The invention claimed is:
1. A fault detection system for detecting faults in a turbine
engine, the system comprising: a plurality of fault detectors the
plurality of fault detectors adapted to receive device data from
the turbine engine and determine a plurality of fault conclusions
from the device data; and a diagnostic aggregation mechanism, the
diagnostic aggregation mechanism adapted to receive the plurality
of fault conclusions and utilize at least one valid multiple fault
set to aggregate the plurality of fault conclusions and isolate a
likely fault in the turbine engine, wherein the diagnostic
aggregation mechanism is adapted to aggregate the plurality of
fault conclusions and isolate a likely fault in the turbine engine
using a hybrid Dezert-Smarandache Theory (DSmT) aggregation
rule.
2. The system of claim 1 wherein the at least one valid multiple
fault set is determined from a systematic model of the turbine
engine, the systematic model describing effects of a plurality of
failure modes on the turbine engine.
3. A fault detection system for detecting faults in a turbine
engine, the system comprising: a plurality of fault detectors, the
plurality of fault detectors adapted to receive device data from
the turbine engine and determine a plurality of fault conclusions
from the device data; and a diagnostic aggregation mechanism, the
diagnostic aggregation mechanism adapted to receive the plurality
of fault conclusions, partition the plurality of fault conclusions
based on dependency of data used by the plurality of fault
detectors to generate the plurality of fault conclusions into high
dependent sets, and further partition high dependent sets into weak
dependent supersets, and to isolate a likely fault in the turbine
engine by aggregating conclusions in the high dependent sets and by
aggregating conclusions in the weak dependent supersets.
4. The system of claim 3 wherein the diagnostic aggregation
mechanism is adapted to aggregate conclusions in the high dependent
sets using a high dependence fusion rule, and wherein the high
dependence fusion rule comprises averaging conclusions based on
high dependent device data.
5. The system of claim 4 wherein the diagnostic aggregation
mechanism is adapted to aggregate conclusions in the weak dependent
supersets using a weak dependence fusion rule, and wherein the weak
dependence fusion rule comprises averaging conclusions using a
weighted average.
6. The system of claim 5 wherein the diagnostic aggregation
mechanism is further adapted to aggregate conclusions based on
independent evidence and the aggregated conclusions in the weak
dependent supersets using an independent evidence rule.
7. The system of claim 6 wherein the independent evidence rule
comprises a hybrid Dezert-Smarandache Theory (DSmT) aggregation
rule.
8. A method of detecting faults in a turbine engine, the method
comprising the steps of: receiving device data from the turbine
engine; determining a plurality of fault conclusions using a
plurality of fault detection techniques, the plurality of fault
detection techniques each using a subset of the device data; and
isolating likely faults by utilizing at least one valid multiple
fault set to aggregate the plurality of fault conclusions and
isolate a likely fault in the turbine engine, wherein the step of
isolating likely faults by utilizing at least one valid multiple
fault set to aggregate the plurality of fault conclusions and
isolate a likely fault in the turbine engine comprising using a
hybrid Dezert-Smarandache Theory (DSmT) aggregation rule.
9. The method of claim 8 wherein the step of isolating likely
faults by utilizing at least one valid multiple fault set to
aggregate the plurality of fault conclusions and isolate a likely
fault in the turbine engine comprises determining the at least one
valid multiple fault set using a systematic model of the turbine
engine, the systematic model describing effects of a plurality of
failure modes on the turbine engine.
10. A method of detecting faults in a turbine engine, the method
comprising the steps of: receiving device data from the turbine
engine; determining a plurality of fault conclusions using a
plurality of fault detection techniques. the plurality of fault
detection techniques each using a subset of the device data;
partitioning the plurality of fault conclusions based on dependency
of data used by the plurality of fault detection techniques to
generate the plurality of fault conclusions into high dependent
sets, and further partitioning high dependent sets into weak
dependent supersets; and isolating likely faults in the turbine
engine by aggregating conclusions in the high dependent sets and by
aggregating conclusions in the weak dependent supersets.
11. The method of claim 10 wherein the aggregating conclusions in
the high dependent sets comprises aggregating using a high
dependence fusion rule, and wherein the high dependence fusion rule
comprises averaging conclusions based on high dependent device
data.
12. The method of claim 11 wherein the aggregating conclusions in
the weak dependent supersets comprises aggregating using a weak
dependence fusion rule, and wherein the weak dependence fusion rule
comprises averaging conclusions using a weighted average.
13. The method of claim 12 wherein the step of isolating likely
faults further comprises the step of aggregating conclusions based
on independent evidence and the aggregated conclusions in the weak
dependent supersets using an independent evidence rule.
14. The method of claim 13 wherein the independent evidence rule
comprises a hybrid Dezert-Smarandache Theory (DSmT) aggregation
rule.
15. A program product comprising: a fault detection program, stored
on a computer-readable recordable medium which, when run, causes a
processor to predict a fault in a turbine engine, the fault
detection program including: a plurality of fault detectors, the
plurality of fault detectors adapted to receive device data from
the turbine engine and determine a plurality of fault conclusions
from the device data; and a diagnostic aggregation mechanism, the
diagnostic aggregation mechanism adapted to receive the plurality
of fault conclusions and utilize at least one valid multiple fault
set to aggregate the plurality of fault conclusions and isolate a
likely fault in the turbine engine, wherein the diagnostic
aggregation mechanism is adapted to aggregate the plurality of
fault conclusions and isolate a likely fault in the turbine engine
using a hybrid Dezert-Smarandache Theory (DSmT) aggregation rule.
Description
FIELD OF THE INVENTION
This invention generally relates to diagnostic systems, and more
specifically relates to prognosis systems for mechanical
systems.
BACKGROUND OF THE INVENTION
Modem mechanical systems can be exceedingly complex. The
complexities of modem mechanical systems have led to increasing
needs for automated prognosis and fault detection systems. These
prognosis and fault detection systems are designed to monitor the
mechanical system in an effort to predict the future performance of
the system and detect potential faults. These systems are designed
to detect these potential faults such that the potential faults can
be addressed before the potential faults lead to failure in the
mechanical system.
One type of mechanical system where prognosis and fault detection
is of particular importance is aircraft systems. In aircraft
systems, prognosis and fault detection can detect potential faults
such that they can be addressed before they result in serious
system failure and possible in-flight shutdowns, take-off aborts,
delays or cancellations. Engines are, of course, a particularly
critical part of the aircraft. As such, fault detection for
aircraft engines are an important part of an aircraft's fault
detection system.
In some applications it is desirable to use multiple fault
detection techniques to monitor a mechanical system. In these
applications the different fault detection techniques can focus on
different part of the system, and can use different data and
algorithms in determining if potential fault exists. One issue in
utilizing multiple fault detection techniques is the ability to
correctly harmonize the multiple potential conclusions derived from
the concurrent use multiple different fault detection techniques.
Specifically, using multiple fault detection techniques can
potentially result in multiple incomplete, ambiguous or
contradictory conclusions. Unfortunately, previous techniques for
combining incomplete conclusions from multiple fault detection
techniques have had limited ability to deal with multiple
concurrent faults and dependent evidence. This has reduced the
ability to utilize multiple different fault detection techniques to
accurately detect potential faults.
Thus, what is needed is an improved system and method for combining
conclusions from multiple fault detection techniques in mechanical
systems such as turbine engines.
BRIEF SUMMARY OF THE INVENTION
The present invention provides a system and method for combining
conclusions from multiple fault detection techniques to isolate
likely faults in a turbine engine. The embodiments of the invention
provide the ability to effectively deal with multiple concurrent
faults in the engine. Additionally, the embodiments of the
invention provide the ability to correctly characterize multiple
conclusions generated from evidence having different levels of
interdependence.
Specifically, the present invention provides a multi-technique,
multi-fault detection system and method that isolates likely
fault(s) in turbine engines. The system receives device data,
including sensor data from the turbine engine. The device data is
passed to a plurality of fault detectors. Each fault detector
analyzes the device data to determine a likelihood of one or more
particular faults in the turbine engine. The plurality of fault
detectors can use a different fault detection technique and
different types and combinations of device data to determine the
likelihood of fault in the turbine engine. Additionally, the
plurality of fault detectors can determine the likelihood of
different types of faults in the turbine engine. Accordingly, each
of the plurality of fault detectors outputs a conclusion that
indicates the likelihood of a corresponding fault or faults in the
turbine engine.
The conclusions from the fault detectors are passed to a diagnostic
aggregation mechanism that isolates likely faults from the
plurality of conclusions. In one embodiment, the diagnostic
aggregation mechanism determines possible multi-fault combinations
that are indicated by the conclusions. This is done by identifying
possible combinations of valid multiple fault sets using a model of
the turbine engine the conclusion range of the plurality of fault
detectors. The valid multiple fault sets can include both static
multiple fault sets that are determined from the turbine engine
model and run-time multiple fault sets determined from the
conclusions themselves. When the valid sets of multiple conclusions
are identified, they can be combined using an aggregation rule.
This allows the diagnostic aggregation mechanism to identify any
combinations of multiple faults that are occurring in the turbine
engine
The diagnostic aggregation mechanism can also aggregate faults
based on the dependency of the data used to generate the
conclusions, e.g., the amount of overlap in the data used by the
various techniques. This allows the diagnostic aggregation
mechanism to give aggregated multiple conclusions based on
independent evidence more weight than those where the conclusions
are based on dependent evidence, as conclusions that are based on
the same evidence are not independent and thus do not have the same
probative value. Thus, by taking into account the degree of
dependency in the data used to generate the conclusions, the
conclusions can be aggregated in way that properly takes into
account their relatively probative values.
In this embodiment, the conclusions based on device data with high
dependency are aggregated using a high dependency aggregation rule,
and the resulting high-dependency sets are then further aggregated
using a weak dependency rule. Finally, any conclusions based on
independent evidence can be aggregated using an independent
combination rule. The resulting aggregation determines which
fault(s) are most likely indicated by the plurality of conclusions,
taken into account the dependency of the device data used to
generate the conclusions. Thus, the embodiments of the invention
provide a system for combining conclusions from multiple fault
detection techniques to isolate likely faults in a turbine
engine.
The foregoing and other objects, features and advantages of the
invention will be apparent from the following more particular
description of a preferred embodiment of the invention, as
illustrated in the accompanying drawings.
BRIEF DESCRIPTION OF DRAWINGS
The preferred exemplary embodiment of the present invention will
hereinafter be described in conjunction with the appended drawings,
where like designations denote like elements, and:
FIG. 1 is a schematic view of a multi-technique multi-fault
detection system;
FIG. 2 is a flow diagram of a method for generating a database of
static valid multiple fault sets;
FIG. 3 is a flow diagram of a method for aggregating conclusions of
valid sets;
FIG. 4 is a flow diagram of a method for aggregating conclusions
based on evidence dependency;
FIG. 5 is a table view of a schematic view of an exemplary turbine
engine fault set and an exemplary diagnostic algorithm set; and
FIG. 6 is a schematic view of a computer system in accordance with
an embodiment of the invention.
DETAILED DESCRIPTION OF THE INVENTION
The present invention provides a system and method for combining
conclusions from multiple fault detection techniques to isolate
likely faults in a turbine engine. The embodiments of the invention
provide the ability to effectively deal with multiple concurrent
faults in the engine. Additionally, the embodiments of the
invention provide the ability to correctly characterize multiple
conclusions generated from evidence having different levels of
interdependence.
The various embodiments of the invention provide multi-technique,
multi-fault detection system and method that isolates likely
fault(s) in turbine engines. Turning now to FIG. 1, a schematic
view of a multi-technique, multi-fault detection system 100 is
illustrated. The system 100 receives device data 110, including
sensor data from the turbine engine. The device data is passed to a
plurality of fault detectors 102. Each fault detector 102 analyzes
the device data to determine a likelihood of one or more particular
faults in the turbine engine. The plurality of fault detectors 102
can use a different fault detection technique and different types
and combinations of device data to determine the likelihood of
fault in the turbine engine. Additionally, the plurality of fault
detectors 102 can determine the likelihood of different types of
faults in the turbine engine. Accordingly, each of the plurality of
fault detectors 102 outputs a conclusion that indicates the
likelihood of a corresponding fault or faults in the turbine
engine.
The conclusions from the fault detectors are passed to a diagnostic
aggregation mechanism 104. The diagnostic aggregation mechanism 104
evaluates the conclusions from the plurality of fault detectors
102, and generates isolated faults 106. The isolated faults 106
indicate the most likely fault or faults indicated by the plurality
of conclusions.
In one embodiment, the diagnostic aggregation mechanism 104
isolates faults by evaluating the plurality of conclusions and
determining possible multi-fault combinations that could be
occurring. This is done by identifying possible combinations of
valid multiple fault sets using a model of the turbine engine and
the conclusion range of the plurality of fault detectors. The valid
multiple fault sets can include both static multiple fault sets
that are determined from the turbine engine model and run-time
multiple fault sets determined from the conclusions themselves.
When the valid sets of multiple conclusions are identified, they
can be combined using an aggregation rule.
In isolating faults, the diagnostic aggregation mechanism 104 can
also evaluate the plurality of conclusions based on the dependency
of the data used to generate the conclusions, e.g., the amount of
overlap in the data used by the various techniques. This allows the
diagnostic aggregation mechanism 104 to give aggregated multiple
conclusions based on independent evidence more weight than those
where the conclusions are based on dependent evidence, as multiple
conclusions that are based on the same or overlapping evidence are
not independent and thus do not have the same probative value.
Thus, by taking into account the degree of dependency in the data
used to generate the conclusions, the diagnostic aggregation
mechanism 104 can effectively aggregate conclusions and isolate
faults.
In one embodiment, diagnostic aggregation mechanism 104 aggregates
conclusions based on device data with high dependency using a high
dependency aggregation rule, and the resulting high-dependency sets
are then further aggregated using a weak dependency rule.
Furthermore, the diagnostic aggregation mechanism 104 aggregates
any conclusions based on independent evidence using an independent
combination rule. The resulting aggregation determines which
fault(s) are most likely indicated by the plurality of conclusions,
taking into account the dependency of the device data used to
generate the conclusions. Thus, the embodiments of the invention
provide a fault detection system 100 for combining conclusions from
multiple fault detection techniques to isolate likely faults in a
turbine engine.
One suitable aggregation rule that can be used by the diagnostic
aggregation mechanism 104 is a hybrid Dezert-Smarandache Theory
(DSmT) aggregation rule. The hybrid DSmT rule allows the diagnostic
aggregation mechanism 104 to identify any potential combinations of
multiple faults that are occurring in the turbine engine.
Additionally, the hybrid DSmT rule can be used to aggregate
conclusions from fault detection techniques based on independent
evidence. In general, hybrid DSmT is a variation of the
Dempster-Shafer (D-S) statistical framework technique. The D-S
statistical framework is a known technique typically used for
diagnostic evidence aggregation. However, in traditional D-S
analysis, evidence from multiple techniques is assumed to result
from independent evidence and a single fault. Attempting to use
traditional D-S techniques in multi-fault applications requires
relaxing the single fault assumption, and significantly increases
the computational complexity. Furthermore, traditional D-S analysis
has limited ability to aggregate conclusions where the conclusions
from the algorithms are not independent. For example, when multiple
fault detection algorithms use the same sensor data. Thus,
traditional D-S techniques cannot be effectively used in
applications such as turbine engines where various different fault
detection techniques use related evidence and multiple faults are
possible.
DSmT based techniques provide the ability to relax the assumption
that elements in the frame of discernment must be mutually
exclusive. In general, DSmT formally combines sources of
information represented in terms of belief functions, and is
typically focused on uncertain, highly conflicting and imprecise
sources of evidence. The foundation of DSmT is the definition of
Dedekind's lattice, also called a hyper-power set of the frame of
discernment, where the frame of discernment is set of all possible
single faults The hyper power set of the frame of discernment
breaks with the classic assumption that elements in a frame of
discernment must be mutually exclusive, and thus can be used to
relax the single fault assumption of traditional D-S techniques. As
will be described in greater detail below, in one embodiment, a
hybrid DSmT rule of combination is used to aggregate conclusions of
valid sets determine any likely multi-fault combinations and thus
isolate the faults occurring in the turbine engine. Additionally,
the hybrid DSmT rule can be used to aggregate conclusions from
fault detection techniques based on independent evidence.
Turning now to FIG. 2, a method 200 for determining multiple
concurrent faults in a turbine engine is illustrated. In general,
method 200 identifies valid multiple fault sets which are used to
reduce the number of non-empty elements in the hyper power set to a
more manageable size. Specifically, by identifying valid multiple
fault sets, the number of non-empty elements in the hyper power set
can be reduced to single faults and potentially valid multiple
fault combinations, excluding all other multiple fault
combinations. The valid multiple fault sets, as part of the reduced
hyper power set, can then be used to aggregate conclusions from
multiple fault detectors and isolate the most likely faults. Each
valid multiple fault set identifies two or more faults that can
accurately identified using the various fault detection techniques
employed by the plurality of fault detectors. For example, each set
can include a listing of faults that can occur simultaneously.
Thus, the valid multiple fault sets each identify multiple faults
that can occur and be detected simultaneously. During operation of
the turbine engine, these valid multiple fault sets can then be
used to determine when detected multiple faults indicate an actual
multiple fault condition, and thus can be used to isolate the most
likely faults that are occurring in the turbine engine.
Specifically, by identifying valid multiple fault sets the
non-empty elements of the hyper power set can be limited to single
fault conclusions and valid multiple fault sets, allowing the
conclusions from multiple fault detection techniques can be
aggregated more efficiently, and a more precise detection of the
most likely faults obtained.
In method 200, a systematic model 202 is used as the basis for
determining the valid multiple fault sets. In general, a systematic
model is a definition of the system under consideration, and is
typically human understandable and machine interpretable. The
systematic model 202 for a turbine engine would typically be
generated from a functional description of the turbine engine and a
failure mode effect and consequence analysis of the engine. This
analysis determines the various different ways that components can
fail what the effects of those failures would be on the larger
system. This analysis serves as the static basis for the systematic
model of the turbine engine. Additionally, the systematic model can
include dynamic functions that identify how various diagnostic
algorithms used by the fault detectors can provide information on
the failure modes of the turbine engine. In this function each
algorithm is correlated to the appropriate components in the
turbine engine for which the algorithm provides diagnostic
information. From this, a database of the various components in the
turbine engine, how those components can fail, how the potential
effects of those components can affect the larger system provides
the static basis for the systematic engine.
A specific example of how a systematic model can be applied to a
turbine engine will now be discussed. In the systematic model, each
component c.sub.i in the turbine engine is associated with one or
more functions g.sub.j. For example, the metering valve component
within the fuel control unit is associated with the function
provide metered fuel. Next, each component c.sub.i in the turbine
engine is associated with one or more failure modes h.sub.k. These
failure modes describe the mechanism of failure of the components.
Additionally, attributes such as mean time to failure, repair cost,
or repair time are correlated to various failure modes. Next, the
systematic model includes how failure modes prevent the various
components from achieving their functions. The systematic model
captures this information in the form of a mapping, e.g., { . . . ,
(g.sub.j, h.sub.k, . . . }. This notation shows that failure mode
h.sub.k affects function g.sub.j, etc.
It should be noted that function g.sub.j associated with component
c.sub.i may also depend on another function g.sub.s. This function
dependency can be captured by making g.sub.j depend on function
g.sub.s. As a general modeling rule, g.sub.s, should not depend on
any other function associated with component c.sub.i. In other
words, all functions associated with component c.sub.i are
independent. Finally, diagnostic algorithms used by the various
fault detectors are associated with functions. This information is
combined into a systematic model that will be used to isolate
faults in the turbine engine. The systematic model for a turbine
engine can be developed using a variety of tools and techniques.
For example, commercial off-the-shelf tools can be used to create
the model from the information described above.
Returning to method 200, the first step 204 is to identify a fault
conclusion range for each algorithm. In this step, the systematic
model 202 is used to determine what the possible valid conclusions
are from each fault detection algorithm used by the various fault
detectors. For example, in a systematic model 202 where nodes
represent failure modes {h.sub.k}, functions {g.sub.j}, and
algorithms {a.sub.i}, the arcs represent functional dependency,
failure dependency, and algorithm location. By analyzing the
systematic model, the frame of discernment for each fault detector
algorithm is determined, as well as what parts of the frame of
discernment are mutually exclusive. This information can then be
used to generate valid sets of multiple faults.
For example, if an algorithm a.sub.l, identifies a fault associated
with function g.sub.j, this implies two things: 1) one or more
failure modes g.sub.j are present; or 2) the functions on which
g.sub.j depend have failed. If g.sub.j depends on functions
g.sub.a, g.sub.b, the same reasoning can be applied to g.sub.a and
g.sub.b, and those failure modes examined, up the backward chain of
dependent functions. This backward chaining can be continued till
all failure nodes are examined. Thus, the algorithm a.sub.i can be
used to partition a failure mode set {h} in to {h.sup.A.sub.l} and
{h}/{h.sup.A.sub.l}, where {h.sup.A.sub.l} denotes the failure
modes that can be implicated by algorithm a.sub.l. The failure
modes implicated by an algorithm are defined as the range
r(A.sub.l) of the algorithm. When such an analysis is performed for
every algorithm using the systematic model 202, a set of all
possible multiple fault sets .THETA. can defined as follows:
.THETA..times..function. ##EQU00001## Where h.sub.0 is an unknown
failure mode.
With the range of each algorithm identified, the next step 206 is
to identify any valid multiple fault sets for each possible
combination of fault detection algorithms. This can be done by
examining each algorithm to determine the valid sets of faults for
each algorithm. Then, the conclusions of different algorithms are
examined to determine combinations of conclusions outside the
intersections of valid sets.
One technique that can be used to identify the valid multiple fault
sets is an iterative examination of the conclusions of the various
algorithms. In this technique, for each pair of fault detection
algorithms, all two-element non-empty sets are determined by taking
a union of the difference in each algorithm range. In each
successive pass, the candidate k-element sets for a set of k
algorithms are generated by joining (k-1) element sets of those
algorithms, and deleting those that include any (k-1) subsets that
were determined to be relatively empty at the previous iteration.
This process is continued until no more non-empty sets can be
generated. From this examination, a database 208 of static valid
multiple fault sets is created.
As will be described in greater detail below with reference to FIG.
3, this database 208 of static multiple fault sets can then be used
to aggregate conclusions from multiple diagnostic algorithms to
determine multi-fault combinations and isolate likely faults.
Turning now to FIG. 3, a method 300 for aggregating conclusions
from multiple fault detectors is illustrated. In general, this
method uses sets of static multiple faults and a hybrid
Dezert-Smarandache Theory (DSmT) rule to isolate likely faults from
a plurality of conclusions generated by a plurality of fault
detectors.
The first step 304 is to identify run-time valid multiple fault
sets from a plurality of conclusions 302 received from the fault
detection system. In one embodiment, the algorithms themselves are
implemented to identify run-time valid multiple sets. Thus, when an
algorithm concludes that there are multiple faults it identifies a
run-time valid set.
With the run-time valid multiple fault sets identified, the next
step 306 is to merge the static valid multiple fault sets generated
in method 200 with the run-time valid multiple fault sets generated
in step 304. The merged valid multiple fault sets are then used
with the hybrid DSmT rule to isolate the most likely conclusions
from the conclusions received from the fault detection system. This
allows the diagnostic aggregation mechanism to identify the likely
combinations of multiple faults that are occurring in the turbine
engine.
The hybrid DSmT based techniques provide the ability to relax the
assumption that elements in the frame of discernment must be
mutually exclusive. In general, DSmT formally combines sources of
information represented in terms of belief functions, and is
typically focused on uncertain, highly conflicting and imprecise
sources of evidence. Thus, in one specific embodiment, a hybrid
DSmT rule of combination is used in step 308 to aggregate
conclusions of valid sets determine any likely multi-fault
combinations and thus isolate the faults occurring in the turbine
engine. This rule preferably uses a hybrid model M(.THETA.) adapted
for use either more than two independent sources of information. As
one example, the rule of combination for the hybrid model
M(.THETA.) can be defined as:
m.sub.M(.THETA.)(F)=.phi.(F)[S.sub.1(F)+S.sub.2(F)+S.sub.3(F)] (2.)
Where .THETA. is the set of conclusions, .phi.(F) is the
characteristic non-emptiness function of a set F, i.e., .phi.(F)=1
if F.noteq.O and .phi.(F)=0 otherwise, where O.ident.O.sub.M, O.
O.sub.M is the set of all elements of the Dedekind's lattice
D.sup..THETA. that have been forced empty through the constraints
of the model M and O is the classical/universal empty set.
S.sub.l(F) corresponds to the free DSmT rule of combination for k
independent sources based on the free DSmT model, and is given
by:
.function..times..times..di-elect
cons..THETA..times..times..times..times..times..times..times..function.
##EQU00002## And S.sub.2(F) represents the mass of all relatively
and absolutely empty sets which are transferred to the total or
relative ignorance and is given by:
.function..times..times..di-elect cons..THETA. .di-elect
cons..times..times..times..times..function. ##EQU00003## And
S.sub.3(F) represents the sum of relatively empty sets to the
non-empty sets
.function..times..times..di-elect cons..THETA.
.times..times..times..times..di-elect
cons..times..times..times..times..function. ##EQU00004## With
U.ident.u(F.sub.l) V u(F.sub.2) . . . u(F.sub.k) where u(F) is the
union of all singletons h.sub.i that compose F and
I.sub.t.ident.h.sub.1h.sub.2 . . . h.sub.n is the total ignorance.
It should be note that in turbine engine applications where the
algorithms do not provide conclusions that satisfy non-existential
constraints rule S.sub.2(F) is typically not applied.
In turbine engine fault detection applications, evidence
aggregation can be viewed as a dynamic fusion problem where the
hybrid model M(.THETA.) changes each time the fault detectors post
results, i.e., some of the elements which were not empty at one
posting may become empty the next and vice versa. In one
embodiment, each time the results of the fault detector algorithms
are entered, the hybrid DSmT rule, based on a new hybrid model
M(.THETA.), is applied.
To determine which elements of the hybrid model M(.THETA.) are
empty the model is examined at both the diagnostic algorithm level
and the knowledge fusion level. Namely, if any one of the fault
detection algorithms determines that a certain set of faults occurs
simultaneously, the corresponding element in the hybrid model
M(.THETA.) is determined to be relatively non-empty. This
determination may not be sufficient in all cases, as some
algorithms may have different levels of expertise and will
therefore have different levels of discernment. This can prevent
the accurate diagnosis of faults falling outside the frames of
discernment. Therefore, when several algorithms each identify
faults outside the intersection of their frame of discernment,
knowledge fusion models can assume that those faults are occurring
simultaneously and will assign appropriate beliefs based on the
hybrid combination rule.
In addition to dealing with multiple fault conclusions, the
embodiment of the invention can be applied to isolate faults from
conclusions based on different levels of dependent evidence. For
example, in cases where some fault detection conclusions are based
on the same sensor data or even the same features in the sensor
data. In this embodiment, the diagnostic aggregation mechanism
aggregates conclusions based on device data with high dependency
using a high dependency aggregation rule, and the resulting
high-dependency sets are then further aggregated using a weak
dependency rule. Furthermore, the diagnostic aggregation mechanism
aggregates any conclusions based on independent evidence using an
independent combination rule. The resulting aggregation determines
which fault(s) are most likely indicated by the plurality of
conclusions, taking into account the dependency of the device data
used to generate the conclusions. This type of aggregation can be
combined with isolation of fault from multiple conclusions, or
performed independently.
In one specific implementation of this method, the fault detection
algorithms are first partitioned into the non-overlapping, high
dependence sets, where each set represents algorithms that base
their conclusions on highly dependent evidence. For example, fault
detection algorithms that use the same sensor features from the
turbine engine are partitioned into the same set high dependence
set. The resulting sets of conclusions can then further partitioned
into non-overlapping, weak dependence sets, where each set
represents conclusions based on weakly dependent evidence. For
example, the conclusions of algorithms that use the same sensors,
but not necessarily the same features in the sensor data are
partitioned into the same weak dependence set.
In one embodiment, this partitioning of the algorithms is based on
a determined proportion of the overlapping evidence, referred to
herein as a degree of dependence w.sub.dep. By determining the
degree of dependence, the conclusions can be consistently
categorized into the weak and high dependency sets, and combined
using the appropriate rule. A variety of different techniques can
be used to determine the degree of dependence. For example, it can
be set apriori to an appropriate value, such as 0.5 for highly
dependent evidence and 0.25 for weekly dependent evidence.
Alternatively, the degree of dependence w.sub.dep can be
dynamically calculated based on some scheme. One method of
calculating the degree of dependence can be based on the degrees of
freedom of the algorithms in each dependence set. For example, if
the algorithms use different methods but the same sensor features,
then the amount of independent evidence can be w.sub.dep=1/3 of the
entire body of evidence, because there are three degrees of freedom
(method, sensor and feature) and the algorithms differ only in the
method used. In such a case, the amount of dependent evidence used
by each high dependence algorithm would be w.sub.dep=2/3.
Finally, the conclusions from each weak dependence set are
aggregated using an independent combination rule, such as a DSmT
combination rule based on independent sources of evidence.
As one illustrative example, consider four algorithms A.sub.1,
A.sub.2, A.sub.3 and A.sub.4. These algorithms apply a variety of
methods to one or more sensor features. Algorithm A.sub.1 applies
method m.sub.1 to feature f.sub.1 of sensor s.sub.1. Algorithm
A.sub.2 applies method m.sub.2 to feature f.sub.1 of sensor
s.sub.1. Algorithm A.sub.3 applies method m.sub.1 to feature
f.sub.2 of sensor s.sub.2. Finally, algorithm A.sub.4 applies
method m.sub.3 to feature f.sub.3 of sensor s.sub.2. The algorithms
are partitioned into three high dependency sets,
HDSet.sub.1={A.sub.1, A.sub.2}, HDSet.sub.2={A.sub.3} and
HDSet.sub.3={A.sub.4}. The conclusions of each such set are then
further partitioned into the following two weak dependence sets:
WDSet.sub.1={HDSet.sub.1} and WDSet.sub.2={HDSet.sub.2,
HDSet.sub.3}. With the conclusions so partitioned, the conclusions
can be aggregated using suitable aggregation rules.
Turning now to FIG. 4, a method 400 for aggregating conclusions
from multiple fault detectors using dependent evidence is
illustrated. In this method, the first step 404 is to aggregate the
plurality of conclusions 402 from techniques with high dependency.
This step uses a high dependence fusion rule, on the high
dependence sets of fault detector conclusions. One example of a
high dependence fusion rule can be defined as:
.function..times..A-inverted..noteq..THETA..times..function..times..A-inv-
erted..noteq..THETA..function..THETA..A-inverted..noteq..THETA..times..fun-
ction. ##EQU00005## Where m.sub.rec,ident reconciled belief from
multiple algorithms using the identical features and sensors, and
where k is the number of algorithms in the high dependency set. In
one embodiment, the high dependence rule operates by taking an
average of all algorithms in the highly dependent set. For example,
if five fault detection algorithms provide 0.1, 0.2, 0.3, 0.4, 0.5
belief to a fault hypothesis of "Broken blade" then m.sub.A1=0.1,
m.sub.A2=0.2, and so on. When the algorithms all use the same
feature of the same sensor, the different algorithms do not provide
any additional support. Thus, the results can be properly
aggregated by averaging the belief of the five algorithms. The
result of step 404 is thus a single, aggregated conclusion for each
set of high dependency algorithms.
With the conclusions for techniques with high dependencies
aggregated, the next step 406 is to further aggregate the
aggregated conclusions from techniques with weak dependency.
Included in the conclusions that are aggregated with this step are
the previously aggregated conclusions for sets of high dependency
algorithms. Thus, step 406 aggregates any single conclusions with
weak dependency with any aggregated conclusions from the previous
step to generate a single aggregated conclusion that defines the
ultimate conclusion from all techniques based on high dependent
evidence and weak dependent evidence. Weak dependency refers to
conclusions that may use different evidences. For example,
algorithms that use related data from different sensors, such as
the same type of sensor but from different manufactures. To
aggregate this data, a weak dependence fusion rule is used. For
example, the weak dependence rule can comprise taking an average
that is weighted to account for weak dependency. As one example of
a weak dependency rule, each algorithm A.sub.i has a range
r(A.sub.i). Each set of faults {F.sub.j,A.sub.i} includes the
conclusions that could be occurring simultaneously. To determine
the weighted average of these conclusions, the weighted average of
the conclusions m(F) can be defined as:
.function..times..A-inverted..times..function..times..A-inverted..di-elec-
t cons..THETA. ##EQU00006## Where the quantity m is an average
belief assigned by F by all algorithms, and where any hypothesis
F.sub.j,A.sub.i can be assigned a belief as:
.function..PHI..function..function..times..times..times..function.
##EQU00007## And modified as:
.function..function..times..E-backward..times..times..times.
##EQU00008##
In equation 10, the conflict between two algorithms that use highly
dependent evidence (i.e., identical sensors and features) is
`distributed` to multiply their respective beliefs to all possible
combinations of F. For example, if algorithm A.sub.1 and algorithm
A.sub.2 use identical sensors and features, but algorithm A.sub.1
claims fault F.sub.1 with a m(A.sub.1) degree of belief, algorithm
A.sub.2 claims fault F.sub.2 with a m(A.sub.2) degree of belief,
then there is a conflict. Equation 10 resolves this conflict by
assigning m(A.sub.1) and m(A.sub.2) to the combined fault F.sub.1
and F.sub.2, or m(A.sub.1^A.sub.2).
The result of step 406 is thus an aggregation of conclusions based
on dependent evidence for each set of weak dependency algorithms.
Thus, steps 404 and 406 have aggregated conclusions from techniques
with high and weak dependency into a single conclusion representing
the ultimate conclusion of conclusions based on any type of
dependent evidence. With the sets of conclusions for techniques
with weak dependencies aggregated, the next step 408 is to further
aggregate the conclusions based on an independent techniques. This
step uses an independent combination rule. For example, the
independent combination rule can use a DSmT combination rule, such
as the rule used in method 300 to aggregate conclusions of valid
sets to determine multi-fault combinations. In this step the DSmT
combination rule is used to aggregate the conclusions of the
independent fault detection techniques based on the dependency of
the evidence to isolate the most likely faults in the turbine
engine. It should be noted that this step can utilize an
independent combination rule such as DSmT because the remaining
conclusions are based on independent evidence. Specifically,
because conclusions based on dependent evidence were separated from
conclusions based on independent evidence, and the conflict between
conclusions based on dependent evidence were resolved in steps 404
and 406 above, the remaining conflict between conclusions can be
resolved using an independent combination rule. Thus, step 408 uses
an independent combination rule to resolve conflicts between the
remaining conclusions based on independent evidence and the
aggregated conclusions from step 406.
The resulting aggregation of step 408 thus determines which
fault(s) are most likely indicated by the plurality of conclusions,
taking into account the dependency of the device data used to
generate the conclusions.
A detailed example of how the techniques described above can be
applied to a turbine engine in an auxiliary power unit (APU) will
now be discussed. In a typical turbine engine, ambient air entering
the engine is split at the plenum. Part of the air enters the main
compressor, and the remaining air enters the load compressor. The
flow of air through the load compressor is regulated by the inlet
guide vanes. High pressure air from the main compressor enters the
combustor, in which fuel is introduced using a series of nozzles.
The air-fuel mixture is burnt continuously and smoothly in the
combustor. Hot combustion gases are expanded in the turbine engine,
which drives the engine shaft. Part of this useful work is expended
by the main compressor, and the remaining part is expended by the
load compressor to provide bleed air and power the gearbox. The
gear box provides power to a generator, which provides electrical
power to the vehicle. Exhaust air from the turbine engine is
introduced combustor using a fuel control unit. The engine consists
of a single hollow shaft supported by a set of oil-cooled bearings.
Lubrication oil is provided using a gear pump and air-cooled heat
exchanger. Fast acting valves provide the necessary surge control
and protection. The turbine engine is started using a battery
operated started motor.
A variety of different fault detection techniques can be used on
such a turbine engine. These fault detection techniques can use a
variety of different algorithms, from simple threshold checking
calculations to complex multivariate model based estimations.
Turning now to FIG. 5, a listing 502 of an exemplary set of
potential turbine engine faults and a listing 504 of an exemplary
set of diagnostic algorithms are illustrated. This listing 502
includes various types of faults that can be predicted for a
turbine engine. Using systematic modeling, a database of static
valid multiple fault sets can be determined. As the set of
conclusions .THETA. can be defined as:
.times..function..times..times..THETA. ##EQU00009##
In this example, algorithm A.sub.1 analyzes the starter motor
current during the startup phase and provides evidence of faults in
the starter motor. During startup, the turbine gear is not yet
engaged, and hence the range of A.sub.1 is very small. That is,
r(A.sub.1)=H.sub.1. On the other hand, algorithm A.sub.2 monitors
the feedback signal from the inlet guide vane position sensor and
is influenced by mechanical problems in the vane mechanism such
that r(A.sub.2)=H.sub.7. This measurement is made when the APU is
delivering a full load. Given range of these algorithms, it is
clear that m.sub.A1, and m.sub.A2 should not be treated as
conflicting evidence, but instead should be treated as capable of
providing evidence supporting a multiple fault hypothesis.
In this example, the calculation of valid multiple fault sets can
be illustrated by algorithms A.sub.3, A.sub.4, and A.sub.5.
Algorithm A.sub.3 calculates pneumatic balance for the APU. Torque
generated by the turbine is roughly equal to the fuel burnt in the
combustor and the torque expended by the load compressor. This
balance is performed when the APU is idling, i.e., under no load
conditions. Algorithm A.sub.4 performs the load balance under
exclusive electrical load, i.e., the APU providing electrical
power. Under these conditions, the guide vanes are fully closed and
the algorithm thus cannot implicate the IGV or the load compressor.
Algorithm A.sub.5 calculates the residual in the heat rejection
system. Heat generated at the bearing and generator is removed by
the oil cooler.
Using a systematic model, the range of the exemplary algorithms in
the exemplary turbine engine is found to be:
r(A.sub.3)={H.sub.1,H.sub.3,H.sub.3}
r(A.sub.4)={H.sub.1,H.sub.3,H.sub.4}
r(A.sub.5)={H.sub.1,H.sub.3,H.sub.4,H.sub.5} (14.)
Algorithms A.sub.6, A.sub.7, A.sub.8, A.sub.9 provide evidence
toward turbine and fuel control faults. However, each of these
algorithms use different methods and work on different sensors.
Algorithm A.sub.6 is the fuel flow jump detector and it applies a
fuzzy logic method to the magnitude feature of the fuel flow
sensor. Algorithm A.sub.7 monitors for changes in the magnitude
feature of the fuel flow sensor, but uses a sequential probability
ratio testing method. Algorithm A.sub.8 is an EGT rate detector,
and uses a slope feature of the exhaust gas temperature sensor,
with the slope being analyzed using a fuzzy logic method. Finally,
algorithm A.sub.9 applies a hypothesis testing method to the
variance feature of the exhaust gas temperature sensor. Given the
overlap between these algorithms with respect to common sensors, it
is needed to explicitly calculate the dependence between evidences
m.sub.A6, m.sub.A7, m.sub.A8, m.sub.A9.
With the dependency of the various algorithms calculated, the
conclusions can be partitioned into high dependency and weak
dependency sets, as described above. Then, with the algorithms
partitioned according the dependence of evidence, the conclusions
from those algorithms can be aggregated using aggregation method
400 described.
The multi-technique, multi fault detection system and method can be
implemented in wide variety of platforms. Turning now to FIG. 6, an
exemplary computer system 50 is illustrated. Computer system 50
illustrates the general features of a computer system that can be
used to implement the invention. Of course, these features are
merely exemplary, and it should be understood that the invention
can be implemented using different types of hardware that can
include more or different features. It should be noted that the
computer system can be implemented in many different environments,
such as onboard an aircraft to provide onboard diagnostics, or on
the ground to provide remote diagnostics. The exemplary computer
system 50 includes a processor 110, an interface 130, a storage
device 190, a bus 170 and a memory 180. In accordance with the
preferred embodiments of the invention, the memory system 50
includes a multi-technique, multi fault detection program.
The processor 110 performs the computation and control functions of
the system 50. The processor 110 may comprise any type of
processor, include single integrated circuits such as a
microprocessor, or may comprise any suitable number of integrated
circuit devices and/or circuit boards working in cooperation to
accomplish the functions of a processing unit. In addition,
processor 110 may comprise multiple processors implemented on
separate systems. In addition, the processor 110 may be part of an
overall vehicle control, navigation, avionics, communication or
diagnostic system. During operation, the processor 110 executes the
programs contained within memory 180 and as such, controls the
general operation of the computer system 50.
Memory 180 can be any type of suitable memory. This would include
the various types of dynamic random access memory (DRAM) such as
SDRAM, the various types of static RAM (SRAM), and the various
types of non-volatile memory (PROM, EPROM, and flash). It should be
understood that memory 180 may be a single type of memory
component, or it may be composed of many different types of memory
components. In addition, the memory 180 and the processor 110 may
be distributed across several different computers that collectively
comprise system 50. For example, a portion of memory 180 may reside
on the vehicle system computer, and another portion may reside on a
ground based diagnostic computer.
The bus 170 serves to transmit programs, data, status and other
information or signals between the various components of system
100. The bus 170 can be any suitable physical or logical means of
connecting computer systems and components. This includes, but is
not limited to, direct hard-wired connections, fiber optics,
infrared and wireless bus technologies.
The interface 130 allows communication to the system 50, and can be
implemented using any suitable method and apparatus. It can include
a network interfaces to communicate to other systems, terminal
interfaces to communicate with technicians, and storage interfaces
to connect to storage apparatuses such as storage device 190.
Storage device 190 can be any suitable type of storage apparatus,
including direct access storage devices such as hard disk drives,
flash systems, floppy disk drives and optical disk drives. As shown
in FIG. 6, storage device 190 can comprise a disc drive device that
uses discs 195 to store data.
In accordance with the preferred embodiments of the invention, the
computer system 50 includes a deterioration prediction program.
Specifically during operation, the deterioration prediction program
is stored in memory 180 and executed by processor 110. When being
executed by the processor 110, fault detection program receives
data from the device being monitored and isolates fault detection
predictions from that data.
As one example implementation, the fault detection prediction
system can operate on data that is acquired from the mechanical
system (e.g., aircraft) and periodically uploaded to an internet
website. The analysis is performed by the web site and the results
are returned back to the technician or other user. Thus, the system
can be implemented as part of a web-based diagnostic and prognostic
system.
It should be understood that while the present invention is
described here in the context of a fully functioning computer
system, those skilled in the art will recognize that the mechanisms
of the present invention are capable of being distributed as a
program product in a variety of forms, and that the present
invention applies equally regardless of the particular type of
computer-readable signal bearing media used to carry out the
distribution. Examples of signal bearing media include: recordable
media such as floppy disks, hard drives, memory cards and optical
disks (e.g., disk 195), and transmission media such as digital and
analog communication links.
The embodiments and examples set forth herein were presented in
order to best explain the present invention and its particular
application and to thereby enable those skilled in the art to make
and use the invention. However, those skilled in the art will
recognize that the foregoing description and examples have been
presented for the purposes of illustration and example only. The
description as set forth is not intended to be exhaustive or to
limit the invention to the precise form disclosed. Many
modifications and variations are possible in light of the above
teaching without departing from the spirit of the forthcoming
claims.
* * * * *